Experiments
How does a self-play agent actually play Star Realms? Each experiment below trains or probes
the policies and reports what they do. Results pages are generated by the
run-experiment command and served statically, so re-reading them never re-runs the
(expensive) training.
Leaderboard
- Elo & win-ratio matrix — how the methods rank head-to-head, the data the training harness prints, on a page. Also reports how the trained agents' games end (decisively, by the turn cap, or by the move cap).
Field notes
- An engineering challenge — why high-authority training is expensive — what happened when we tried to train deep at 50 authority, why it is not an engine bug, and how we bounded the cost.
Opening-buy studies
A running question: with a fixed turn-1 trade row (Junkyard, Fleet HQ, Blob World, Battle Pod, Trade Pod), which card does the agent buy — and how does that depend on the authority the game starts with and on how long the agent trained?
- Experiment 1 — Strategy vs. starting authority — does the agent shift from Battle Pod (aggression) to Trade Pod (economy) as authority rises?
- Experiment 2 — The under-trained model anchors — a lightly-trained model buys the card it was trained on, whatever authority it now plays.
- Experiment 3 — The well-trained model adapts — a heavily-trained model buys the correct card for the authority in play, however it trained.